Claude Code v2.1.145: JSON Agent Listing, OTEL Span Fixes, Security Patch, and More

Claude Code v2.1.145 ships with a handful of practical additions and a security fix you'll want to update for. The headline feature is claude agents --json, which lets you script against live agent sessions — think tmux-resurrect, status bars, or custom session pickers.
JSON Agent Listing
Run claude agents --json to get a machine-readable list of active Claude sessions. Each entry includes agent ID, status, and other metadata. This makes it straightforward to build integrations without parsing terminal output.
Security Fix: Permission Bypass
A permission-prompt bypass was fixed where bare variable assignments to non-allowlisted environment variables in Bash commands were auto-approved. Update to ensure your env vars are still gated by the permission system.
OTEL Span Improvements
Telemetry gets smarter: claude_code.tool spans now carry agent_id and parent_agent_id attributes. Background subagent spans properly nest under the dispatching Agent tool span, giving you accurate trace trees.
Status Line & GitHub PR Info
The status line JSON input now includes detected GitHub repo and PR information — useful for automated workflows that react to the current repository context.
Plugin Discover & Browse
Before installing a plugin, /plugin Discover and Browse screens now show the plugin's commands, agents, skills, hooks, and MCP/LSP servers. No more guesswork.
Terminal Tab Title
The claude agents terminal tab title now displays the awaiting-input count, so you can alt-tab and see at a glance which agent needs attention.
Bug Fixes & Behavior Changes
- Read tool: When a whole-file read exceeds the token limit, the tool now returns a truncated first page with a "PARTIAL view" notice instead of a hard error.
- MCP slash commands: Raw server validation errors are replaced with a message naming the missing argument and showing expected usage.
- Spinner freeze: The spinner and elapsed-time display no longer freeze after terminal resize or refocus.
- Windows PowerShell: Cross-project resume hint now uses
;as the command separator (fixes Windows PowerShell 5.1). - Task list order: Tasks created at once now render in deterministic order.
- Marketplace banner: Stale "Failed to install" banner no longer shows when already installed.
- PR badge: Updates immediately after
gh pr createand similar PR-state-changing commands. - Non-ASCII names: Agent Teams teammates with non-ASCII names no longer fail API calls due to invalid header encoding.
/review: No longer uses the deprecatedprojectCardsGraphQL query — fixes errors on repos with Classic Projects.- Plugin validate: Now flags
skills:entries pointing to a file instead of a directory, with a suggestion to use the parent directory. - Skill context:fork: Fixed an infinite loop where a skill using
context: forkcould repeatedly re-invoke itself.
Update to v2.1.145 to get the security fix, scripting capability, and smoother MCP/skill behavior.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Anthropic Policy Update Bans Third-Party Tools for Claude Pro/Max Users
Anthropic updated their policy in February 2026 to explicitly ban any script, wrapper, or third-party tool usage with Claude Pro or Max plans, resulting in lifetime bans for users who violate this policy. High-tier Max plan users engaging in heavy coding sessions are being targeted in a March 2026 enforcement wave.

Claude Opus 4.6's effort=low parameter differs from other providers' low-reasoning modes
Claude Opus 4.6's effort=low parameter controls general behavioral effort, not just reasoning depth, unlike OpenAI's reasoning.effort=low or Gemini's thinking_level=low. This caused agents to make fewer tool calls, be less thorough in cross-referencing, and ignore parts of system prompts about web research.

Opus 4.6 excels at research, Gemini 3.1 Pro has better judgment in forecasting benchmark
A benchmark of 1,417 binary forecasting questions separates research and judgment performance: Claude Opus 4.6 leads in agentic research, Gemini 3.1 Pro wins on fixed-evidence calibration. GPT-5.4 and Grok 4.20 show little change between conditions.

Research shows AI users often accept LLM answers without verification
University of Pennsylvania research found AI users engage in 'cognitive surrender,' accepting LLM answers with minimal scrutiny. In experiments, users accepted correct AI answers 93% of the time and incorrect answers 80% of the time, even when AI was wrong half the time.